New opportunities for high-resolution countrywide tree type mapping

Transcription

1 New opportunities for high-resolution countrywide tree type mapping Lars T. Waser, Bronwyn Price, Nataliia Rehush, Marius Rüetschi, and David Small* Swiss National Forest Inventory Swiss Federal Research Institute WSL, Birmensdorf, Switzerland *Remote Sensing Laboratories (RSL), University of Zurich 1

2 Table of Contents Table of Contents Background, user demands, state-of-the-art Countrywide tree type mapping - Remote sensing and training data - Approaches, products, challenges Conclusions 2

3 Background Increasing demand on countrywide tree species information Forest management Forest industry Renewable energy sources etc. Beyond forestry sector: biodiversity, nature conservation etc. *Fassnacht, F. E., Latifi, H., Stereńczak, K., Modzelewska, A., Lefsky, M., Waser, L. T., (2016). Review of studies on tree species classification from remotely sensed data. Remote Sensing of Environment, 186,

4 Background Why do we spatially estimate forest attributes? Lack of spatial information beyond National Forest Inventory (NFI) sample plots* -> Spatial products needed for operational NFI applications NFI plots - Tree types Tree species Volume Health etc. *Barrett, F., McRoberts, R.E., Tomppo, E., Cienciala, E., and Waser, L.T., A questionnaire-based review of the operational use of remotely sensed data by NFIs. Remote Sensing of Environment 174:

5 Background 5 Is countrywide mapping of tree types feasible?

6 What is needed by the user? User demands (high) expectations on tree species maps What input data are available (remote sensing data, reference data) Continuity of these data sets (regular updating?) What level of detail? (e.g. single tree level, plot, stand level) In the last 40 years, advances in remote sensing technologies (new sensors, 3D point clouds, machine learning etc.) However, (only) recently from case study to countrywide level 6

7 State-of-the-art Building a bridge Gap between research and practice: optimal conditions versus operational constraints => difficult to implement Research, ISI journals User, Forest practice (NFIs) UNESCO, world heritage Parc Ela 7

8 Table of Contents Table of Contents Background, user demands, state-of-the-art Countrywide tree type mapping - Remote sensing and training data - Approaches, products, challenges Conclusions 8

9 Input data Remote sensing data of Switzerland ADS40/80/100 sensor with cm RGBI aerial imagery updated every 3 years (since 2005) by Swiss Federal Office of Topography LiDAR with ~ points/m2 ( ), since 2017 full-waveform Sentinel-1 (SAR) / Sentinel-2 (1C, 8 bands), m spatial resolution GSD 10cm GSD 25cm 9

10 Training / reference data Input data From Swiss National Forest Inventory two-phases sample based survey, continuously visited (9 year circle) on a 1.4 km regular grid Aerial stereo-image interpretation Terrestrial survey From individual field mapping / image interpretation 10

11 Countrywide tree type mapping Tree type map of Switzerland ( km 2 ) 11

12 Countrywide tree type mapping Tree type map of Switzerland ( km 2 ) Distinction of broadleaved / coniferous trees at 3 m spatial resolution Input: RGBI ADS80 aerial images, remote sensing indices, digital terrain model from ALS data Training data: Digitized tree polygons Highly automated workflow using Random Forest (RF) in R Model accuracies: %, Kappa (5 *10-fold CV) => Overestimation of coniferous tree fraction *Waser, L.T.; Ginzler, C.; Rehush, N., Wall-to-Wall Tree Type Mapping from Countrywide Airborne Remote Sensing Surveys. Remote Sensing, 9, 766 *Waser, L.T. et al., Remote Sensing, 6, *Waser, L.T. et al., Remote Sensing of Environment, 115,

13 Countrywide tree type mapping Challenges ADS80 (August 2015) Topography: Steep terrain (shaded crowns) Phenology: Date of image acquisition => Overestimation of coniferous trees Broadleaved Coniferous

14 14Countrywide tree type mapping Improvements Improved error estimations using ensemble modelling (RF, SVM, Logistic regression, ANN, knn) => partly satisfactory and new opportunities Use of multitemporal Sentinel-2 (multispectral) data to minimize problems due to phenology, steep terrain => partly satisfactory

15 Countrywide tree type mapping Sentinel-2 time series ADS80 (8/2015), 3m Sentinel-2 (8/2016 & 2/2017), 10m

16 16Countrywide tree type mapping Improvements Improved error estimations using ensemble modelling (RF, SVM, Logistic regression, ANN, knn) => partly satisfactory and new opportunities Use of multitemporal Sentinel-2 (multispectral) data to minimize problems due to phenology, steep terrain => partly satisfactory Sentinel-1 SAR winter / summer data Backscatter signals (VV, VH) from SAR Usage of DTM: slope classes, aspect classes

17 Countrywide tree type mapping Combination of Sentinel- 1 and 2 time series Sentinel-1 SAR (2016 / 2017) Sentinel-1 and -2 (2016 / 2017) => Feasibility to seperate larches from other conifers

18 Countrywide tree type mapping Accuracy assessment Model accuracies > 96%, Kappa (5 *10-fold CV) Map accuracies - Stereo-image Interpreted Areas (IAs) with 25 points - Agreement in predicted (tree type map) and observed (NFI) broadleaved fractions in the IAs => Averaged out, no general underestimation of broadleaved tree fraction 18

19 Countrywide tree type mapping 19 Map accuracies

20 Countrywide tree type mapping Table of Contents Background, user demands, state-of-the-art Countrywide tree type mapping - Remote sensing and training data - Approaches, products, challenges Conclusions 20

21 Conclusions Is countrywide mapping of tree types feasible? YES 21

22 Conclusions Countrywide tree type mapping feasible (3-10 m spatial resolution, high accuracies) Restrictions (topography, shadows, phenology) minimized (Combination of Sentinel-1 / 2 time series) Remaining icreasing demand on countrywide products Providing spatial explicit information which is not given by NFI plots Ongoing / future research: High temporal / spectral resolution of upcoming sensors Deep learning (CNN) to improve classifications Focus on more tree species 22

23 New opportunities for high-resolution countrywide tree type mapping Lars T. Waser, Bronwyn Price, Nataliia Rehush, Marius Rüetschi, and David Small Thank you for your attention! 23